Dear Jan,
>
> I would like to run the gpu versions of bedpostx (6.5) and eddy_cuda6.5 on the university server which is running CentOS 6 and cuda 6.5 but I get the following errors:
>
> error running bedpostx_gpu:
> ERROR: fit_PVM_single_c_kernel: invalid device function
>
> error running eddy_cuda6.5:
> ...................Allocated GPU # 43941408...................
> CUDA error after call to EddyGpuUtils::InitGpu
> Error message: CUDA driver version is insufficient for CUDA runtime version
my, rather lay, understanding is as follows. A CUDA installation has two parts. One is the runtime environment, and this is what you can have several versions installed of: 6.5, 7.5, 8.0 etc. There is also a driver, and for the driver there is only a single version installed at any given time. The driver is backwards compatible, so if you have a driver associated with for example 9.0 that will work together with CUDA 9.0 and all older versions that you may have installed. But of course the driver is not forward compatible, which means that if you have an old driver it will not work with a newer version of CUDA.
Your error message indicates that the driver that is installed on your system is older than that associated with CUDA 6.5. I don’t know about the first error message, but I think until proven otherwise I would assume it has the same root.
My recommendation would be to install the very latest version of the driver. If that doesn’t solve the problem I would try to redo the entire CUDA installation.
Jesper
>
> I'm able to run both these processes (7.5 versions) on the same data on another computer running ubuntu 14.04 and cuda 7.5. The second error looks like it's related to the cuda version but I don't know about the first one. Any ideas on what may be causing these errors?
>
> Additionally, would it be possible to make the source code available?
>
> Thanks,
>
> Jan
|